2084

2084 explores the intersection of technology, ethics, and the future, with a focus on AI, its applications, and philosophical implications. It delves into AI advancement, music and image generation, business strategies, academia, future transportation, and power sources. The Substack analyzes current trends and theorizes on future developments in AI and societal structures.

Artificial Intelligence Technology and Future Trends Music and Image Generation AI Business and Economics Academia and Industry Interface Ethics and Philosophy Transportation Innovations Power and Energy

The hottest Substack posts of 2084

And their main takeaways

2084: Is ChatGPT self-aware?

98 implied HN points • 14 Dec 22
  1. Philosophical questions arise when considering if AI like ChatGPT is self-aware.
  2. Comparing human sentience to AI behavior raises questions about self-awareness.
  3. It is challenging to determine if AI like ChatGPT is truly self-aware based on external behaviors.

2084: Musing on Music AI

58 implied HN points • 29 Nov 22
  1. Music AI generation is complex due to the large amount of data and correlations needed for accurate music production.
  2. Symbolic music generation using MIDI files or notes is a workaround for the complexity of raw audio data generation.
  3. Current AI music generation tools are still largely research-based and not as advanced as other AI applications like text to image generation.

2084: more understandable text in your text to image images

19 implied HN points • 06 Sep 23
  1. Text-to-image generation can be improved by making text more understandable in images
  2. An approach like Gligen uses bounding boxes and associated text to generate better images
  3. Using OCR with CLIP and Gligen models could enhance the process of injecting text into images

2084: VC dry powder and AI

39 implied HN points • 13 Jan 23
  1. Venture capitalists have a lot of uninvested capital, particularly in AI.
  2. Traditional VC sectors are not performing well, making AI the most attractive sector for investment.
  3. 2023 is a great time to start an AI startup due to high interest and funding available in the sector.

2084: Bacon, Materialism and AI

39 implied HN points • 02 Jan 23
  1. Francis Bacon believed the value of philosophy should be based on usefulness, not perfection.
  2. Bacon was the first genuine materialist, emphasizing improving material conditions as important as moral character.
  3. While Bacon's philosophy aligns with AI development's focus on practicality, ethical considerations are needed for responsible AI advancement.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

2084: Cold Diffusion

39 implied HN points • 26 Dec 22
  1. Arbitrary image transforms can result in good information recovery
  2. Neural networks might be learning the loss of information itself
  3. Diffusion models could have more general applications beyond expectations

2084: Chatting with the future: Redux

39 implied HN points • 09 Dec 22
  1. ChatGPT by OpenAI is impressively good at generating text and responses.
  2. ChatGPT has a large working memory and can hold a lot of context during conversations.
  3. There is potential for future applications of ChatGPT beyond text, such as in audio and video generation.

2084: The Future of Subscriber Email: Introducing Newsreader: AI-Powered Breaking News Email Generation

39 implied HN points • 06 Dec 22
  1. The author created an AI-powered email generator called Newsreader that summarizes breaking news articles from various websites into email updates.
  2. By utilizing GPT-3's capabilities, the author was able to efficiently summarize complex articles in a clear and understandable way.
  3. The project showcases the potential of AI in automating tasks traditionally done through programming, hinting at a future where AI prompts might replace conventional coding methods.

2084: How does Image Generation AI work?

39 implied HN points • 05 Dec 22
  1. Image Generation AI models use diffusion models to remove noise from images and generate new images from noise patterns.
  2. AI models like Dall-E 2, Midjourney, and Stable Diffusion utilize different methods to produce images based on captions and training data.
  3. The future of art and creativity might be transformed by AI, making tasks quicker and more accessible, potentially changing the role of artists.

2084: Chatting with the future

39 implied HN points • 03 Dec 22
  1. New AI models like OpenAI GPT-3 and Meta's AI Cicero show advancements in lifelike conversation and complex game playing.
  2. Future AI tools may become more user-friendly and powerful through platforms like DeepGram and HuggingSpaces.
  3. There is a need for more user-friendly libraries of AI models to simplify setup steps and improve accessibility.

2084: Making Models Faster

39 implied HN points • 02 Dec 22
  1. The team is working on making convolutional neural network models faster by using sparse convolutional layers.
  2. They are designing an architecture that involves a 'input-stable' approach and a unique method of computing convolutions.
  3. To support different operations, they plan to extend their multipliers to Multiplier Accumulators controlled by external bits in a 'tile-flexible' paradigm.

2084: War and Pitch

39 implied HN points • 01 Dec 22
  1. The author discusses Tolstoy's struggle between realistic depictions and moral parables in 'War and Peace.'
  2. The pitch AI model is structured using CM-Bert to label audio and transcripts with investment information.
  3. Using AI and data scraping, the model aims to evaluate pitch videos for investment potential and explore ways to provide feedback for improvement.

2084: Making Stable Diffusion Smaller & Better Pitches Today!

39 implied HN points • 28 Nov 22
  1. Models like Stable Diffusion are impressive but can be slow due to their size.
  2. Techniques like learning rate rewinding can help make models smaller and faster.
  3. Using intermediate activations for training subnetworks can speed up inference time significantly.

2084: Academia and Business

39 implied HN points • 26 Nov 22
  1. Academia and business often seem disconnected, with valuable research not being implemented in practical applications.
  2. Proprietary journals in academia can be seen as profiting off researchers and limiting access to knowledge.
  3. Efforts like HuggingFaces and initiatives funding practical research show promise in bridging the gap between academia and industry.

2084: Work

39 implied HN points • 23 Nov 22
  1. Procrastination is a natural rhythm of the mind and should not always be fought against.
  2. Establishing a strict schedule can be challenging as external events can disrupt it easily.
  3. Ultimately, the only way to overcome work is by actually doing the work.

2084: Neuradisconnect

39 implied HN points • 23 Nov 22
  1. Brain computer interfaces like Neuralink show promise for future technologies like brain-operated vehicles and prosthetics.
  2. Challenges with brain interfaces include the brain's natural defense mechanisms against foreign objects and the complexity of information transfer.
  3. Future questions arise on the practicality and effectiveness of brain-controlled systems compared to AI alternatives.

2084: Power in the future

39 implied HN points • 21 Nov 22
  1. Electrification is crucial for modern society and impacts various systems like water, gas, and food.
  2. While renewables are growing, coal and gas power will still be necessary due to demand and supply challenges.
  3. In the future, adaptations to climate change will be common, with shifts in living environments, but human populations are likely to remain stable.

2084: Open AI GPT-3

39 implied HN points • 20 Nov 22
  1. The capabilities of OpenAI's GPT-3 are vast, handling various tasks with high accuracy.
  2. Future advancements in AI like GPT-4 are expected to be even more powerful and versatile.
  3. AI technologies like GPT-3 have the potential to revolutionize industries, professions and daily tasks by providing efficient and intelligent solutions.

2084: Cars

39 implied HN points • 19 Nov 22
  1. Electric cars need to be more innovative to be truly efficient compared to gasoline cars.
  2. Society often combines past designs with future technology, creating a sense of 'path dependence'.
  3. Aptera Motors is changing the game by designing lighter, more efficient cars using aerodynamics and carbon fiber.

2084: Business

39 implied HN points • 18 Nov 22
  1. Serve your customers well and focus on providing products that meet their needs.
  2. Prioritize quality and efficiency in your business operations.
  3. In the future, there may be a shift towards more localized and community-focused business practices.

2084: MarcRandbot: Speech Synthesis with Mamba

3 HN points • 03 Jan 24
  1. Dataset creation is a crucial part of training models for speech synthesis.
  2. Using the SpeechTokenizer model for creating tokens combining semantic and acoustic information.
  3. Mamba performed better than Transformers for speech synthesis tasks, showing efficiency and quality.

2084: Why can Sydney and ChatGPT think?

19 implied HN points • 20 Feb 23
  1. Transformer models like GPT-3 can generalize tasks well beyond their training data, showing impressive abilities
  2. These models create internal models to perform tasks, indicating a level of reasoning and adaptability
  3. Sydney and ChatGPT may be capable of forming complex internal models, suggesting potential for deeper understanding and thought

2084: Leveraging the Power of Language Models for Long Term Planning

19 implied HN points • 10 Feb 23
  1. Large language models like GPT-3 can do long term planning in complex tasks.
  2. Smaller language models like GPT-J can outperform larger models by additional training on specific tasks like using tools or answering questions.
  3. There is a growing focus on applying language models in practical ways, indicating a promising future for AI-powered solutions.

2084: The AI Wars are starting

19 implied HN points • 09 Feb 23
  1. Google announced Bard to compete with ChatGPT, but Microsoft leads in devops.
  2. Microsoft's new Bing with GPT integration shows promise.
  3. Quora's Poe model aggregator may become a significant player in advanced models.

2084: More Music Generating AI

19 implied HN points • 20 Jan 23
  1. Msanii is a new AI music waveform generator using diffusion models.
  2. It converts Mel spectrograms to audio using a neural network for crisper audio.
  3. Msanii was trained on the Pop909 dataset, producing high-quality music demos.

2084: AI Generated Images with Control

19 implied HN points • 19 Jan 23
  1. GliGEN is a diffusion model that allows control over the output of AI generated images.
  2. It uses an annotated dataset and a gated transformer network to control image composition.
  3. The implementation results in shorter training time and potential for commercial use.

2084: Editing the Facts

19 implied HN points • 16 Jan 23
  1. It's possible to edit and change facts stored in language models by making localized edits to specific groups of neurons.
  2. ROME identifies important neurons for a fact by running a language model twice and adding noise to find the most different neurons.
  3. MEMIT is a mass editor that can change multiple facts at once by editing a range of layers in a language model.

2084: Speech Synthesis in the Future

19 implied HN points • 14 Jan 23
  1. Microsoft has introduced Vall-E, a speech synthesizer that can generate speech in different tones and emotions.
  2. Vall-E learns a numerical codec of audios to generate speech based on a given text and input sound.
  3. Although Vall-E's generated speech may sound somewhat alien, with more data and advancements, it could lead to more lifelike audio in the future.

2084:Music Generating AI

19 implied HN points • 04 Jan 23
  1. Music Transformer uses relative position attention for long distance correlation in music generation.
  2. SaShiMi is a waveform generator based on state space models, offering convincing music output.
  3. State space models allow for computing long distance correlations in music generation.

2084: Positional Encodings on the Outputs of Transformers

19 implied HN points • 04 Jan 23
  1. The idea of adding positional encodings to the output of transformers is being explored.
  2. Positional encodings in transformers can help capture relative positions in sequences.
  3. Including positional factors in transformer outputs could improve long-distance correlations.

2084: Things to look out for in the new year

19 implied HN points • 01 Jan 23
  1. The year 2023 will focus on refining AI models like OpenAK and generative AI art.
  2. GPT-3 enabled features will become prevalent in various applications due to text's universal nature.
  3. Anticipate advancements in multi modality technologies like Text to Everything with the saved funds for 2023.

2084: Away for New Year

19 implied HN points • 30 Dec 22
  1. The post is about being away for New Year's.
  2. There's a recommended video to watch on Transformers.
  3. Readers are encouraged to subscribe for more posts.

2084: Some videos

19 implied HN points • 29 Dec 22
  1. The post recommends watching interesting videos on AI.
  2. The content creator is traveling today and shares the videos for viewers with spare time.
  3. Readers can subscribe for free to receive new posts and support the creator's work.

2084: Ideas on a personal roboassistent - KITT in real life

19 implied HN points • 28 Dec 22
  1. Creating a personal robo assistant for a car using AI models like Speech to Text, Language Processing, and Text to Speech.
  2. Planning to implement basic personality traits and commands for the AI assistant to make it more personable.
  3. Considering setting up an Arduino board with WiFi, speaker, and microphone in the car to have an interactive AI assistant, similar to KITT from Knight Rider.

2084: The end of search

19 implied HN points • 23 Dec 22
  1. Google's reliance on ads and e-commerce in search could be threatened by AI like ChatGPT
  2. ChatGPT's ability to provide instant, accurate answers may make Google less relevant
  3. Google's failure to innovate could lead to a decline in relevance and its eventual replacement by companies like OpenAI

2084: Sketch to CAD, Sketch to Pose, and Image to Sketch

19 implied HN points • 19 Dec 22
  1. New tools can turn sketches into useful data
  2. Advancements are being made in practical segments of industry
  3. AI tools like Sketch2Pose provide helpful solutions for tasks such as mapping sketches to 3D poses

2084: Expert Diffusion - eDiff

19 implied HN points • 13 Dec 22
  1. Using different specially trained diffusion models at each denoising step can lead to better results in image generation.
  2. An ensemble of encoders in eDiff allows for different types of input to be considered at different stages, increasing stability in diffusion models.
  3. eDiff is a step towards making diffusion models more usable for everyday business tasks and art generation.

2084: Variational Autoencoders and where ChatGPT gets its data.

19 implied HN points • 12 Dec 22
  1. Variational Autoencoders are a type of neural network used to encode data with a bottleneck structure
  2. The data used to train ChatGPT includes Common Crawl, WebText2, Books1, Books2, and Wikipedia
  3. Potential future improvements for AI models like GPT-3 may involve dynamically grabbing training data from the internet

2084: AI editing of documents and an automatic narrator

19 implied HN points • 11 Dec 22
  1. New transformer models can process arbitrary documents beyond text
  2. Transformers can generate and edit documents naturally by encoding layout information
  3. Large generative models are becoming more versatile, capable of tasks like auto narration and physics-based animation

2084: AI Hardware - how will the hardware for AI look like in the future?

19 implied HN points • 10 Dec 22
  1. AI hardware for the future will be much larger and more powerful, with models like GPT-4 potentially having 100 trillion parameters.
  2. Companies like Nvidia and Cerebras are investing heavily in AI hardware, with new supercomputers and GPUs being developed to handle the demands of large models.
  3. There is a significant increase in funding going into AI hardware development, leading to advancements in neural network accelerators and custom chips specifically designed for machine learning applications.